The Shift to Expensive Tech and How Custom Chips Could Save the Day
In today’s fast-paced world of artificial intelligence, costs are climbing higher than ever before. What used to be an affordable tool for businesses and researchers might soon become a luxury only big companies can afford. This could create bigger gaps in who gets to use AI effectively, boosting productivity for some while leaving others behind. But there’s hope on the horizon: specialized custom chips, known as ASICs, might step in to replace traditional GPUs in certain AI tasks, potentially making things more efficient and affordable over time.
Recent data shows AI expenses are rising due to demands for powerful hardware, massive energy use, and limited supplies. However, breakthroughs in technology could help balance this out. It looks like access to advanced AI will become more uneven, with large tech firms holding the advantage, but custom solutions offer a path to greater efficiency.
- Training costs for top AI models have jumped about 2.4 times each year since 2016, hitting around $192 million per model in 2024—and projections suggest over $1 billion by 2027 if the pattern continues.
Source: IEEE Spectrum, State of AI 2025 - The cost of running AI models (inference) has dropped dramatically, but bottlenecks in chips and power might turn this around by 2026-2028.
- Custom ASICs like Google’s TPUs can be 10-100 times more energy-efficient than GPUs for specific jobs, which could lower long-term costs but requires significant upfront spending.
Source: arXiv preprint on TPU performance - Data centers’ energy consumption is exploding, from 76 terawatt-hours in 2018 to potentially 580 terawatt-hours by 2028 in the US, putting pressure on power grids and increasing bills.
Source: Grid Strategies Power Demand Forecasts - This trend might widen social divides: small businesses and academics could struggle with access, while big tech controls 80-90% of computing power.
Breaking Down AI Cost Trends
AI involves two main costs: 1) training, which builds the models, and 2) inference, which runs them in real-world applications. Training the most advanced models now costs tens or hundreds of millions of dollars, largely because they need enormous amounts of computing power. According to reports from academic sources, these expenses are growing rapidly.
For instance, the cost to train a leading model like GPT-4 was estimated at $78 million in 2023, but by 2024, similar efforts reached nearly $192 million. Looking ahead, if growth continues at this rate, we could see billion-dollar training runs by 2027. On the brighter side, the cost per operation has improved thanks to better efficiency.
Source: Machine Learning Model Training Cost Statistics 2025
Inference costs have fallen exponentially—from about $0.01 per token in 2018 to $0.00007 per token in 2025—but supply issues could cause a rebound soon. Energy use per token is also a key factor, with data centers becoming more power-hungry.
Hardware and Energy: The Big Cost Drivers
Specialized chips like GPUs are the backbone of AI today, but shortages in advanced components are driving prices up. High-bandwidth memory and packaging technologies are in tight supply, with lead times stretching to months. This adds to the overall cost per computing operation.
Energy is another major hurdle. Data centers now use about 4% of US electricity, and demand is set to double by 2030. Efficiency measures, like power usage effectiveness ratings around 1.4, help a bit, but grid delays and rising electricity prices are making operations more expensive.
Source: Pew Research on US Data Centers Energy Use
Custom Chips: A Potential Game-Changer
Custom ASICs, such as Google’s TPUs, are designed specifically for AI tasks and offer big advantages in energy efficiency and performance for things like running models. They could start replacing GPUs in inference workloads, where speed and low power matter most.
Evidence from technical analyses shows large companies investing in these chips, with commitments to manufacturing capacity. This could lead to cost savings of 20-50% in the long run.
Source: ASIC Chip Market Trends Report
The Social Impact: Who Gets Left Behind?
As costs rise, smaller players like startups and universities might find it harder to compete. Big tech firms already dominate 80% of AI investments, creating an uneven playing field. This could mean wider productivity gaps across industries and regions.
Looking Ahead: Cost Forecasts and Warnings
By integrating hardware, energy, and market data, forecasts suggest AI costs per token could rise 20% by 2030 under current trends. But with partial adoption of ASICs, they might stay stable or even drop 15%. Key signs to watch include chip lead times and energy backlogs.
To counter rising costs, improvements in algorithms and software could offset 30-50% of the growth, especially when paired with custom chips.
Table 1: Evolution of floating-point operation costs
| Year | Training Cost ($M) | $/Token Inference | Energy per Token (kWh) | FLOP/$ (log scale) |
|---|---|---|---|---|
| 2018 | 0.5 | 0.01 | 0.00005 | 10 |
| 2020 | 4 | 0.001 | 0.00003 | 12 |
| 2022 | 20 | 0.0002 | 0.00002 | 14 |
| 2024 | 192 | 0.00007 | 0.000015 | 16 |
| 2025 | 300 (est.) | 0.00005 | 0.00002 | 16.5 |
| 2030 | 10,000+ | 0.0001 (potential rise) | 0.00005 | 17 (plateau) |
Scenarios: How Custom ASICs Affect AI Costs for Everyday Users
Custom ASICs could transform AI costs in several ways, making it more accessible or affordable for end customers like businesses and consumers.
- Boost in Efficiency: These chips could cut inference costs by 20-50%, dropping prices from $0.00005 per token to as low as $0.00002. This would mainly benefit users through cloud services, where lower energy use translates to cheaper subscriptions.
Source: TD Securities AI Infrastructure Report - Market Transformation: If ASICs take over from GPUs, overall AI prices might fall 15-30% by 2030. However, large companies would likely see the benefits first, with savings trickling down to users over time.
- Limited Access Risks: High development costs mean ASICs might stay with big players initially, potentially raising prices for public AI services by 10-20% in the short term due to market concentration.
Conclusion: Europe’s Race to Stay in the Game
As Europe grapples with soaring energy demands and regulatory hurdles, the once-affordable world of AI is teetering on the edge of exclusivity. With data centres already devouring 3.1% of the continent’s power and projections pointing to a 70% surge in AI-related electricity use by 2030, the shift from “cheap” to “expensive” AI could deepen divides in innovation and productivity. Yet, custom chips like ASICs, bolstered by the EU Chips Act, offer a lifeline, promising efficiency gains that might keep Europe competitive against global giants. Here’s what Europeans need to know as this tech transformation unfolds.
- Energy Crunch Hits Home:
Europe’s data centres consumed 96 TWh in 2024—equivalent to the annual electricity use of several small countries—and AI could drive a 70% increase by 2030, straining grids from Frankfurt to Dublin. With EU efficiency targets demanding an 11.7% cut in energy use overall, unchecked AI growth risks higher bills and blackouts, prompting calls for stricter Big Tech regulations. - Regulatory Roadblocks Add to Costs:
The EU AI Act, while aiming to safeguard rights, is set to inflate development expenses by up to 17%, potentially costing the economy over €30 billion by 2025. Recent delays and dilutions pushing high-risk AI rules to 2027 signal a balancing act between innovation and oversight, but critics warn it could slow Europe’s AI adoption compared to deregulated markets elsewhere. - Custom Chips: Europe’s Efficiency Edge?:
Forecasts show Europe’s AI chipsets market exploding at 27% CAGR, with custom ASICs leading the charge under the EU Chips Act. These specialised processors could slash inference costs by 20-50% and boost energy efficiency 10-100 times over GPUs, helping firms from Paris to Berlin compete, but only if investments ramp up to match the projected €337 billion AI market by 2032. - Widening Gaps in Access and Growth:
With hyperscalers controlling 80-90% of compute, smaller EU firms and researchers face barriers, exacerbating productivity divides across the bloc. A Gini-like inequality index could hit 0.7 if costs rise unchecked, yet algorithmic tweaks and ASICs might offset 30-50% of hikes, fostering more inclusive growth if shared ecosystems emerge. - Outlook: Caution Amid Optimism:
Under GPU dominance, there’s a 50-70% chance of token costs rising by 2030, but ASIC adoption potentially capturing 30% of inference by 2028 could stabilise or cut prices 15%. For Europe, this means prioritizing sustainable tech to avoid a “haves and have-nots” future; as ECB estimates suggest AI could double productivity to 2.4% annually, the stakes for balanced policy are high.
Will Europe harness custom tech to democratise AI, or let costs create new divides?
Key Citations
- arXiv: The rising costs of training frontier AI models
- LBNL: 2024 US Data Center Energy Report (Note: Updated for 2025 context with similar projections)
- McKinsey: The cost of compute
- arXiv: Deep Neural Network Approximation for Custom Hardware
- Nature: Densing law of LLMs
- Stanford HAI: 2025 AI Index Report
- MERICS: AI Stack Report
- SCSP: 2025 Gaps Analysis Report
- In focus: Data centres – an energy-hungry challenge
- Energy demand from AI – IEA
- AI: Five charts that put data-centre energy use – and emissions
- Majority of Europeans polled want rules to limit new data centres
- Europe Data Centre Power Demand | ICIS
- Clarifying the costs for the EU’s AI Act – CEPS
- Europe loosens reins on AI – and US takes them off – The Guardian
- EU eases AI, privacy rules as critics warn of caving to Big Tech
- The transformative power of AI: Europe’s moment to act
- Custom SoC (ASIC) Market Overview 2030
- Europe AI Chipsets Market Size & Share, 2033
- Europe Artificial Intelligence Market Size, Share | Forecast [2032]
- State of AI 2025: five key charts for Europeans
- The cost of compute: A $7 trillion race to scale data centers
